Token Cost Radar

Token Cost Radar

November 13, 2026

The token-cost story on November 13 is about the gap between cheaper inference and profitable AI. Capgemini reports that 59% of surveyed large enterprises exceeded AI spending expectations, while Google launched Gemini 4 Hydrogen and a multi-model orchestrator that it says used 24% less quota on a coding benchmark. Baidu added a token marketplace and published aggressive pricing for ERNIE 5.5 Flash Preview. Atos warns that an inexpensive agent pilot can become a very expensive production workflow. The thread connecting them is a move from optimizing token prices toward verifying the value of completed work.

Top Developments (Last 24 Hours)

1Can routing between models cut an agent's bill without reducing task success?

Google introduced Gemini 4 Hydrogen on November 12 at $1.50 per million input tokens and $7.50 per million output tokens across a one-million-token context window. Its new Orchestrator in Antigravity switches between models such as Carbon and Hydrogen within a coding session, assigning expensive planning to a stronger model and execution or verification to a faster one. Google reports near-Carbon success on Terminal-Bench 4.0 with 24% less quota consumed and 19% faster completion on average. These are vendor-reported benchmark results; quota savings are not automatically identical to dollar savings. Real economics depend on model rates, reasoning effort, tool calls, and successful completion.

Google ↗

2Why did 59% of surveyed enterprises overshoot their AI budgets?

Capgemini's November 12 research surveyed 500 senior executives at organizations with annual revenue exceeding $1 billion in ten countries. It found that 59% reported AI spending significantly above expectations or budgets during the preceding year, averaging 24% over plan among those affected. Only 56% expressed confidence in forecasting future costs, and only about one-third could confidently estimate indirect expenses such as human oversight and rework. Nearly three in five intend to increase open-weight model use. These are self-reported survey findings, not audited invoices. The research argues that budgeting must connect tokens, model selection, rework, and safely delivered outcomes rather than relying only on aggregate expenditure.

Capgemini ↗

3Baidu puts model tokens on an exchange and prices cached input at RMB 0.04 per million

At Baidu World on November 12, Baidu introduced Token Exchange, a marketplace intended to let customers compare and buy model services with pricing and performance information. Baidu says the marketplace has more than 100 model resources and averages over 50 million calls and 700 billion tokens daily after its first month. It also introduced ERNIE 5.5 Flash Preview with a one-million-token context window and published API rates of RMB 1 per million input tokens, RMB 4 per million output tokens, and RMB 0.04 per million cached input tokens. Baidu says the weights will be released by November 25. These are company claims and listed prices, not independent quality or reliability measurements. The news underscores how provider selection and cache eligibility can materially change effective inference costs.

Baidu announcement via HPCwire ↗

4Atos asks whether the AI pilot still makes financial sense at production scale

In a November 12 essay, Atos architect Peter Kalmijn shows how a seemingly modest 5,000-token business transaction becomes 50 billion tokens at ten million transactions and 500 billion at one hundred million. Those are illustrative calculations, not production customer data. He recommends measuring the full agent loop, including context carried across calls, tool use, retries, and differences between easy and difficult requests. He also asks whether established business rules should replace repeated model reasoning in some steps. The point is architectural: a low per-request price cannot rescue an unnecessarily expensive workflow multiplied across enterprise volumes.

Atos ↗

From Token Prices to Verified Outcomes

Tokenmaxxing is the practice of maximizing AI token consumption, sometimes confusing usage volume with productivity. Tokenminimizing is the practice of eliminating unnecessary tokens while preserving useful results. Token yield measures completed, useful work relative to tokens consumed. The progression from tokenmaxxing through tokenminimizing toward token yield has a clear financial hook in this week's enterprise spending research: lower unit prices matter, but accepted outcomes, retries, and human correction determine the real return. Searches did not establish meaningful fresh adoption of tokenminning, tokenmining, or modelmaxxing as distinct expressions.

Google Decisions API

On November 11, Google introduced a specialized API for bounded decisions such as tool selection, policy checks, and model routing. The company says Gemini Decision 1 evaluates as many as 1,000 questions against shared context of up to 256,000 tokens at $0.10 per million input tokens with no billed output tokens. It returns scores and probabilities rather than generating conventional answer text. Google reports 150 to 250 milliseconds end-to-end latency in its described conditions. The opportunity is to reserve expensive generative reasoning for questions that actually need it, while auditing classification quality and fallback costs.

Google ↗

VentureBeat

VentureBeat's November 12 pricing breakdown cautions that Gemini 4 Hydrogen is not the cheapest lightweight model despite its efficient-workflow positioning. Its standard rates are $1.50 per million uncached input tokens and $7.50 per million output tokens, while eligible Batch and Flex tiers are priced at half the standard rates with different latency and service characteristics. Cached input has a separate lower rate. The practical comparison is not simply Hydrogen versus a cheaper model: it is task success and end-to-end expenditure under the relevant execution tier, including reasoning tokens and retries.

VentureBeat ↗

Microsoft Foundry

Microsoft's July engineering research remains relevant to growing agent tool catalogs. Tool search exposes a small discovery-and-dispatch interface instead of loading all MCP, API, and agent schemas into every prompt. In its reported benchmark, token use fell more than 60% with 50 available tools and more than 97% with 1,000 tools compared with loading the entire catalog upfront. These are benchmark-specific token measurements, not guaranteed savings per completed task. The benefit comes from keeping unused definitions out of context; production systems still need reliable retrieval and authorization.

Microsoft Command Line ↗

Capgemini AI economics framework

Capgemini's earlier research frames AI spending around verified results rather than consumption alone, an approach made timely by its November 12 budget survey. Its framework accounts for direct inference charges as well as human oversight, rework, and governance. The distinction matters for cost routing: an inexpensive model may lower a request's token bill while increasing correction or escalation costs. The framework is a proposed management approach, not a universal accounting standard, and its usefulness depends on assigning credible values and costs to accepted outcomes.

Capgemini Research ↗

Research Watch

FlowPlan: Reducing the Token Cost of Coding Agents with Abstraction-Guided Planning and Verification

Submitted November 9, FlowPlan uses function-level control-flow or data-flow abstractions to plan and verify code repairs instead of repeatedly rereading raw source. Across six evaluated benchmark-model settings it reduced token use without a statistically significant change in issue-resolution rate. In Claude Haiku 4.5 experiments, consumption was 32% to 35% of baseline on SWE-bench Verified and 46% to 47% on SWE-bench Pro. The results are specific to those benchmarks and agents.

Why it matters: Repeated exploration and post-edit rereading can dominate coding-agent context. Structured verification offers a way to reduce work without simply denying the agent information it needs.

arXiv ↗

Think or Remember: The Exchange Rate of Agentic Inference

Submitted November 7, this paper models the cost of keeping or discarding previous reasoning in long-running agents. On recorded Qwen3-8B ScienceWorld runs, discarding earlier reasoning cost 1.74 times as much as keeping it when the server reused generated states, but only 0.34 times as much when nothing was reused. Live reruns also found accuracy losses of 9 to 23 percentage points on selected HotpotQA settings when reasoning history was removed. The outcomes depend on cache behavior and model configuration.

Why it matters: Deleting context is not automatically economical. The cost of remembering depends on what the inference server actually caches and whether lost context triggers additional or worse reasoning.

arXiv ↗

The SIMPLE Law for Energy Cost of LLM Inference

Submitted November 2, this work derives an analytical model for inference energy across attention architectures. On traces from 13 dense models running on NVIDIA H100 GPUs, the researchers report 11.97% mean absolute percentage error without per-model calibration. They find that longer reasoning budgets can produce superlinear energy growth, while prefix caching removes much of prefill energy but not the full cost of managing long-lived attention state. These are measured and modeled energy findings, not a universal API-pricing rule.

Why it matters: A token has a published price, but its physical production cost depends on context length, cache state, model architecture, and reasoning budget. Energy models expose the difference between billing units and actual computation.

arXiv ↗

Phrase of the Day

“AI Yield”

AI Yield is the verified business value of AI-assisted outcomes minus the full cost of delivering them safely, including oversight, rework, and governance. Capgemini uses this term in its enterprise AI economics framework, and its November 12 survey reports that only about one in five organizations calculates it.

  1. AI adoption
  2. Tokenmaxxing
  3. Tokenminimizing
  4. Token discipline
  5. Token yield
  6. Cost per successful task
  7. AI Yield
  8. Verified business value

AI Yield extends the token conversation from engineering efficiency to financial returns. A team can reduce token use without improving output, or increase token use while creating enough value to justify it. Capgemini's survey shows why this distinction matters: budget overruns are widespread, but many executives cannot confidently estimate indirect costs or tie spending to verified outcomes. Google illustrates one optimization path by routing execution to less expensive configurations, while Baidu's published prices illustrate how sharply cached and uncached token costs can differ. Atos raises the architectural question of whether each model call is necessary at all. None of those interventions guarantees higher business value until the completed task, correction effort, and total cost are measured together. The practical progression runs from counting consumption, to avoiding unnecessary consumption, to measuring useful work, and finally to establishing whether that work produces a defensible economic return.

A smaller token bill is progress. A larger verified return is the point.

Capgemini Research Institute ↗

The jCodeMunch read

Today's research on coding-agent exploration and the broader debate over AI economics point to the same principle: avoid processing material that does not help finish the task. jCodeMunch reduces code-reading tokens by more than 95% in its measured code-retrieval benchmark through tree-sitter symbol retrieval and byte-precise context. That is one specific contribution to lower workflow cost, not a substitute for measuring successful completion, retries, and the business value of the result.

See how the 96.3% per-lookup figure is measured →

← All editions